Episode termination

end_criterion is a TaskSpec field (with a --end-criterion override on examples/train.py) that says when the episode ends:

end_criterion	terminates when	use
"failure" (default)	g < 0 (+ timeout). Never on reach.	reach deeper — the agent keeps going after reaching, so the reach-avoid value climbs with l up to the g ceiling
"reach-avoid"	g < 0 or (g ≥ 0 and l ≥ 0)	reach and stop — the episode ends at the target boundary
"timeout"	only the env timeout	diagnostic / pure value-learning

This is the environment half of a pairing whose algorithm half is the learner's terminal_type (safety_sb3 §4). They are orthogonal; all pairings are constructible. The pairing that learns to reach deeper into the target is end_criterion="failure" + terminal_type="all".

How it is implemented

Implemented as a mjlab DoneTerm (zoo_reach_success, fires on g ≥ 0 ∧ l ≥ 0) added only in reach-avoid mode — a real termination term, so mjlab auto-resets on the same step rather than one step late. Default "failure" adds no term.

Defaults reproduce prior behavior exactly: an audit of all safety tasks found none currently terminates on success, so every task stays "failure" and is bit-identical. Switch a task to reach-and-stop by setting end_criterion="reach-avoid" on its TaskSpec, or per-run with --end-criterion.

See also

• Margins — the g / l the criterion tests.
• The task registry — the TaskSpec field.
• The car-goal tutorial — car_goal uses end_criterion="reach-avoid" (drive there and stop).